Last month Google revealed that it's making its own custom artificial intelligence processing chips. Now it turns out that you may already be using them.
The chips are called TPUs, short for Tensor Processing Unit, named for Google's open source AI framework, TensorFlow. Today at the 2016 WIRED Business Conference in New York City, the head of Google's cloud business, Diane Greene, said that Google is already using the chips. So far, the company is using them for AlphaGo, its search ranking algorithm RankBrain, Street View, Photos and speech recognition. "(The TPU) has been in our data centers for over a year now," she said.
Google's cloud customers can also tap into the power of the TPU chips, if they're running applications that can benefit from the chips, Greene said. Google claims TPUs can handle machine learning tasks far more quickly than graphics processors or other chips.
The revelation underscores the importance of both AI and cloud computing to Google's future. As WIRED's Cade Metz pointed out during the onstage interview, at least one Google executive has said that cloud computing services could become a bigger business for Google than advertising. But the competition is steep: Amazon is still the leader in the cloud computing market, and more traditional enterprise services companies such as IBM and Microsoft have their own offerings. These rivals also enjoy longstanding relationships with the sorts of large businesses that cloud computing companies are desperate to sign up.
If Google really does need those business relationships to succeed, Greene joked that the company could always just acquire Oracle. But her real idea is to focus on engineering and innovation rather than sales in order to grow. The company is already putting its machine learning know-how to use on its cloud through services like its Cloud Prediction offering. But Google's custom chips take this potential competitive advantage to the next level. Every cloud computing company offers an enormous amount of computing power and storage. Only Google has TPUs.
Correction 12:00 AM ET 6/17/2016: An earlier version of this story said that TPUs are in use in more than 100 projects, including Android's voice recognition system, the "Smart Reply" feature of its Inbox app, and Google's new cross-application search service Springboard—in short. Greene was actually referring the TensorFlow, not TPUs.
